SkyPilot v0.8.0 adds zero-egress Hugging Face storage across AWS, GCP, Azure

SkyPilot, the open-source multi-cloud orchestrator from UC Berkeley, now mounts Hugging Face Hub datasets and model checkpoints directly into training jobs on AWS, GCP, or Azure without incurring cloud-provider egress charges. Users define a storage mount in YAML pointing to any public or private HF repo, and SkyPilot handles authentication and streaming. The data remains on Hugging Face's infrastructure and flows to compute instances on demand — a 40 GB dataset that would cost $3.60 in AWS transfer fees alone when downloaded to GCP now routes through Hugging Face's CDN at zero cost.

The feature decouples compute from storage, letting teams train on spot instances wherever capacity is cheapest, checkpoint to HF Hub, and avoid the multi-hundred-dollar egress bills that typically lock workflows to a single cloud. SkyPilot's config syntax now accepts store: huggingface alongside S3, GCS, and Azure Blob, with automatic retry logic for transient errors. A five-line YAML example in the Hugging Face blog mounts a 40 GB dataset for distributed training across three clouds without duplicating data—a workflow that would cost hundreds in egress on a traditional cloud-only stack. The change landed in v0.8.0 on July 7, 2026. SkyPilot is Apache 2.0 licensed and runs on any Kubernetes cluster or bare VM.